The runtime is healthy for the selected home directory.
OpenClawBrain verify
Check that it is running and doing its job.
After install, status --detailed is the fast check. It tells you whether OpenClawBrain is loaded and serving on the home you chose.
Use proof only when you want a saved bundle you can keep or share later.
Healthy reference
These four signals are what matter. If your host shows all four, the runtime is working. Pack ids and counters vary by host; the four lines below are the fast read.
The runtime is proven loaded, not just installed on disk.
The runtime is serving a promoted memory pack.
The memory system is active.
STATUS ok
loadProof=status_probe_ready
serve state=serving_active_pack
routeFn available=yesReference output from the healthy exercised host.
Current release
The exercised host, the published package, and the repo are three separate things. Keeping them separate is what makes verification honest.
OpenClawBrain 0.4.45
The current public release. It keeps the same install and proof flow, but makes plain status lighter by default while keeping status --detailed as the deeper live check.
Nonzero truth across all four signals
The exercised host shows STATUS ok, proves that the memory layer is loaded, shows that it is serving, and shows that the install is in sync.
Boundary: this page tells you whether the system is loaded and working on your host. It is not a promise of perfect answers.
Fresh-host note: a fresh install may show
STATUS warn even when the runtime signals are healthy. That is expected. If one installed part updated and another did not, OpenClawBrain should say that clearly instead of silently passing. Use status --detailed as the primary check, not proof alone.
Latest measured improvement
The latest published proof work is bounded and specific, but it is also a real win. On a reviewed frozen test set, OpenClawBrain started catching the cases where memory should help, while still staying disciplined about when not to fire.
Reviewed “memory should help” cases moved from 0 / 10 to 10 / 10
That is the clearest current win. The same checked bundle also stayed clean on the restraint side: unnecessary activations 0 / 65, must-not-fire failures 0 / 69, broad-live regressions 0 / 403.
The product is easier to inspect
The proof lane now makes it easier to see what changed and what the system actually did, instead of burying the result in internal-only details.
Review boundary: some of the newest internal review tooling is still review-only. It is there to make changes easier to inspect before they affect the live system.
Claim boundary: this is a real measured improvement on checked bundles. It is not a claim that learned routing is already universally better or broadly proven online.
When to use each command
Use status --detailed for the fast check. Use proof when you need a durable record.
status --detailedis the fast live check. It tells you whether the memory layer is loaded and whether the install is clean.proofcaptures a durable bundle:summary.md,verdict.json,steps.json, logs, and runtime-load proof in one directory.bundle verdict: success_and_provenmeans install, runtime, load proof, and detailed status all aligned in the captured bundle.- Pack ids and counters depend on host state. The four healthy signals matter more than exact numbers.
- If status already looks healthy, proof should confirm the same story and save it as a bundle.
What we can honestly claim today: OpenClawBrain is installable, inspectable, and working on real hosts.
Run the same check on your host
Keep the home directory the same, restart once, then compare status before you run proof.
Install, restart, status, proof
npx @openclawbrain/[email protected] install --openclaw-home '~/.openclaw'
openclaw gateway restart
npx @openclawbrain/[email protected] status --openclaw-home '~/.openclaw' --detailed
npx @openclawbrain/[email protected] proof --openclaw-home '~/.openclaw'Status first, proof second
If status --detailed already shows the four healthy signals, the proof bundle should confirm and package them. A fresh host may still emit warnings in the proof output.
Go deeper
Install for the commands, how-it-works for the mechanism, troubleshooting when things do not line up.